Letter to the Editor: Measuring what matters in AI-assisted bronchoscopic diagnosis

HIGHLIGHTS Diagnostic positivity reflects acquisition success but does not equate to diagnostic verification or clinical value AI-enhanced navigation improves process metrics without necessarily reducing false-negative malignancy risk Meaningful evaluation requires endpoints linking biopsy yield to diagnostic accuracy and management decisions Composite outcomes may better capture the net clinical benefit of AI-assisted bronchoscopic technologies Dear Editor, We read with great interest the randomized controlled trial by Zhang and colleagues evaluating a deep learning–enhanced virtual bronchoscopic navigation system (SARS-pro) for biopsy of peripheral pulmonary lesions[1]. The authors are to be commended for integrating algorithmic innovation with a prospective randomized design and for demonstrating a substantial increase in diagnostic positivity compared with conventional virtual bronchoscopic navigation. As navigation-assisted bronchoscopy moves toward smaller, more peripheral targets, the field needs precisely this kind of translational engineering work. This study complies with the TITAN 2025 framework for transparent and ethical reporting of AI-assisted research[2]. Beyond the technical advance itself, the study brings into focus a question that now sits at the center of procedural diagnostics: what should count as meaningful progress. Diagnostic positivity is an appealing primary endpoint because it is immediate, measurable, and clinically intuitive. Yet, its interpretability becomes limited when it is asked to stand in for diagnostic truth and clinical value. Positivity is fundamentally an acquisition endpoint, not a verification endpoint. In early lung cancer detection, the decisive issue is not whether a biopsy yields any specific pathological label, but whether diagnostic testing reliably closes post-test malignancy risk in a way that alters management[3]. Diagnostic positivity reflects successful tissue acquisition, but it does not, on its own, distinguish improved lesion targeting from intensified sampling or altered operator behavior prompted by increased navigational confidence. Nor does it quantify the residual probability of missed malignancy after a negative result, which is often the most clinically consequential determinant of downstream care. This distinction becomes more pronounced when positivity aggregates neoplastic and non-neoplastic diagnoses. A navigation system that increases the frequency of obtaining specific benign pathologies may reduce diagnostic ambiguity, yet it may also inflate apparent performance without demonstrating cancer-focused gains. For clinicians and health systems considering adoption, the relevant counterfactual is whether improved positivity translates into fewer missed cancers, fewer repeat procedures, and more timely definitive therapy[4]. Such inferences remain difficult without uniform adjudication of negative results against a reference standard, such as surgical pathology when available or protocolized longitudinal surveillance with predefined re-biopsy criteria. In the absence of this anchoring, higher positivity may represent a redistribution of diagnostic categories rather than a reduction in false reassurance. More broadly, this trial illustrates a recurring challenge in evaluating AI-enabled procedural technologies[5]. Technical advances often first improve process metrics, while both benefit and harm accrue downstream. The most consequential failures of navigation-assisted biopsy are not limited to inability to reach a target, but include confident navigation to an incorrect microanatomical corridor followed by non-representative sampling that delays definitive diagnosis. Establishing net clinical benefit therefore requires endpoints that bind acquisition success to diagnostic correctness and actionability, including malignancy-specific sensitivity, false-negative rates, avoidance of repeat procedures, and management-altering yield. Composite endpoints integrating diagnostic accuracy with downstream clinical impact would better align evaluation with real-world clinical decision-making. These considerations do not diminish the importance of Zhang and colleagues’ contribution. If deep learning–enhanced airway reconstruction can consistently extend navigability into the final centimeters that determine yield, it may represent a genuine step change. The opportunity now is to match this technical progress with outcome definitions capable of supporting confident implementation, reimbursement, and guideline integration. As artificial intelligence continues to reshape procedural diagnostics, the goal should be not only to navigate farther, but to render diagnoses more definitive and care pathways more timely. This study is an important step, and it underscores the need for equal rigor in how success is defined.